packages install
In [1]:
!pip install albumentations==0.4.6
In [2]:
%cd '/content/drive/MyDrive/Colab Notebooks/Plant_Pathology'
In [3]:
from glob import glob
from tqdm.notebook import tqdm
import matplotlib.pyplot as plt
import cv2
# for path in tqdm(glob('data/train_images_l/*.jpg')):
# p_name = path.rsplit('/', 1)[1]
# img = plt.imread(path)
# img = cv2.resize(img, (224,224))
# plt.imsave(f'data/train_images/{p_name}', img)
# # # plt.imsave(f'data/train_images/{p_name}',
# # # plt.imread(path)
# # # )
# # # print(path)
In [4]:
import os
import cv2
import random
import warnings
import torchvision
import numpy as np
import pandas as pd
import seaborn as sns
from PIL import Image
from tqdm.notebook import tqdm
import matplotlib.pyplot as plt
import plotly.graph_objects as go
from insectAug import InsectAugmentation
from torch.utils.data import Dataset, DataLoader
import albumentations as A
from albumentations import (
Compose, OneOf, Resize, Normalize,
# color augmentaion:
RandomBrightness, RandomContrast,
RandomBrightnessContrast, ToGray, HueSaturationValue,
# wether augmentation:
RandomFog,
RandomShadow, RandomRain, RandomSunFlare,
# geomatric transforms:
HorizontalFlip, VerticalFlip, RandomRotate90,
# Blur:
Blur, MotionBlur, MedianBlur,
# noise and cut:
Cutout,
CoarseDropout,
MultiplicativeNoise, GaussNoise, CLAHE, ISONoise, MultiplicativeNoise
)
from albumentations.pytorch.transforms import ToTensorV2
warnings.filterwarnings('ignore')
Let's have a look on the plant pathology raw dataset🤓¶
We can see that the we working with multi-label dataset, each row may has couple of labels
In [5]:
df = pd.read_csv("data/train.csv")
display(df.head(10))
distribution of the labels 📊¶
there are 6 main labels and six more combinations of those labels.
Furthermore, the data is unbalanced and we had to solve it
In [6]:
print(df['labels'].value_counts())
In [7]:
labels_hist = sns.barplot(df.labels.value_counts().index,
df.labels.value_counts())
for item in labels_hist.get_xticklabels():
item.set_rotation(70)
In [8]:
fig = go.Figure(data=[go.Pie(labels=df['labels'].value_counts().index,
values=df['labels'].value_counts().values)])
fig.update_layout(title='Label distribution')
fig.show()
Stratefied split basically preserves the class distribution. When we look at how StratifiedKFold splits the dataset, we saw that it only accepts a one-dimensional target variable, whereas we are trying to pass a target variable with dimensions [n_samples, n_classes]. To deal with this, we perform stratified sampling on the each of the unique label bins.
In [9]:
all_labels = {
'scab':[],
'healthy':[],
'frog_eye_leaf_spot':[],
'rust':[],
'complex':[],
'powdery_mildew':[],
'scab frog_eye_leaf_spot':[],
'scab frog_eye_leaf_spot complex':[],
'frog_eye_leaf_spot complex':[],
'rust frog_eye_leaf_spot':[],
'rust complex':[],
'powdery_mildew complex':[]
}
for image_data, label_data in df.values:
for label in all_labels.keys():
if label_data == label:
all_labels[label_data].append(image_data)
train_df = pd.DataFrame()
val_df = pd.DataFrame()
for label, paths in all_labels.items():
np.random.shuffle(paths)
train_len = round(len(paths)*0.8)
# val_len = len(paths) - train_len
train_paths = paths[:train_len]
val_paths = paths[train_len:]
print(f'** {label} total appearances: {len(paths)}')
print('\ttrain split: ', len(train_paths))
print('\tvalidition split: ', len(val_paths))
for path in train_paths:
tmp_list = [path, label]
train_df = train_df.append(pd.Series(tmp_list, index= df.columns), ignore_index=True)
for path in val_paths:
tmp_list = [path, label]
val_df = val_df.append(pd.Series(tmp_list, index= df.columns), ignore_index=True)
## Save the splitting dataset for future working
# train_df.to_csv(f'data/stratifiedkfold_train.csv', index=False)
# val_df.to_csv(f'data/stratifiedkfold_val.csv', index=False)
Change labels to categorical indices¶
In [10]:
#load train/val datasets
train_df = pd.read_csv("data/stratifiedkfold_train.csv", index_col=False)
val_df = pd.read_csv("data/stratifiedkfold_val.csv", index_col=False)
#shuffle
train_df = train_df.sample(frac=1)
val_df = val_df.sample(frac=1)
Train distribution¶
In [11]:
train_df['labels'].value_counts()
Out[11]:
In [12]:
labels_hist = sns.barplot(train_df.labels.value_counts().index,
train_df.labels.value_counts())
for item in labels_hist.get_xticklabels():
item.set_rotation(70)
In [13]:
fig = go.Figure(data=[go.Pie(labels=train_df['labels'].value_counts().index,
values=train_df['labels'].value_counts().values)])
fig.update_layout(title='Label distribution')
fig.show()
Validition distribution¶
In [14]:
val_df['labels'].value_counts()
Out[14]:
In [15]:
labels_hist = sns.barplot(val_df.labels.value_counts().index,
val_df.labels.value_counts())
for item in labels_hist.get_xticklabels():
item.set_rotation(70)
In [16]:
fig = go.Figure(data=[go.Pie(labels=val_df['labels'].value_counts().index,
values=val_df['labels'].value_counts().values)])
fig.update_layout(title='Label distribution')
fig.show()
Data Frame Encoding¶
In [17]:
# Backup original data
train_df_cp = train_df.copy()
val_df_cp = train_df.copy()
In [18]:
# Convert labels from long string to labels list
train_df_cp['label_list'] = train_df_cp['labels'].str.split(' ')
val_df_cp['label_list'] = val_df_cp['labels'].str.split(' ')
In [19]:
def lbl_lgc(col,lbl_list):
if col in lbl_list:
res = 1
else:
res = 0
return res
In [20]:
lbls = ['healthy','complex','rust','frog_eye_leaf_spot','powdery_mildew','scab']
# create a separate column for each label
for x in lbls:
train_df_cp[x]=0
val_df_cp[x]=0
# fill one's at the relevant position
for x in lbls:
train_df_cp[x] = np.vectorize(lbl_lgc)(x,train_df_cp['label_list'])
val_df_cp[x] = np.vectorize(lbl_lgc)(x,val_df_cp['label_list'])
In [ ]:
train_df_cp.head()
Out[ ]:
In [ ]:
val_df_cp.head()
Out[ ]:
With this finall step we finished preprocceing the raw data🤮
Finally we can see some images!¶
In [22]:
def display_images(train_path,labels,rows,cols):
for i in range(rows):
fig = plt.figure(figsize=(20, 40))
df = list(train_df.loc[train_df['labels'] == labels[i]]['image'])
fontsize=25
label = str(labels[i])
if len(label) > 15:
label = label.replace(' ', '\n')
fontsize=15
try:
label = label.replace('_', ' ')
except:
pass
cat_images = []
for j in range(cols):
img_path = df[j]
img = cv2.imread(os.path.join(train_path,img_path))
img = cv2.resize(img, (512, 512))
cat_images.append(img)
stack_images = cv2.hconcat(cat_images)
image = cv2.cvtColor(stack_images, cv2.COLOR_BGR2RGB)
plt.imshow(image)
plt.axis('off')
plt.title(label, fontsize=fontsize)
plt.show()
# idx += 1
In [23]:
train_image_path = 'data/train_images'
labels_arr = train_df['labels'].unique()
display_images(train_image_path, labels_arr, 12, 4)
The Plant Dataset Class¶
In [ ]:
class Plant_Dataset(Dataset):
def __init__(self, folder_path, data_df, transforms=None, train=True):
self.folder_path = folder_path
self.data_df = data_df
self.transforms = transforms
self.train = train
def __getitem__(self, idx):
img_path = os.path.join(self.folder_path, self.data_df.iloc[idx, 0])
image = cv2.imread(img_path)
if self.train: # generate multi label vector for train/val ground truth
vector = [0]*6
values = self.data_df.loc[self.data_df['image'] == self.data_df.iloc[idx, 0]].values
for i in range(3,9):
num = values[0][i]
vector[i-3] = num
vector = np.asarray(vector)
if self.transforms:
augmented = self.transforms(image=image)
image = augmented['image']
if self.train:
return image, vector # Train\Validation data
else:
return image,self.data_df.iloc[idx, 0] # Test data
def __len__(self):
return len(self.data_df)
In [ ]:
def normalize8(I):
mn = I.min()
mx = I.max()
mx -= mn
I = ((I - mn)/mx) * 255
return I.astype(np.uint8)
def show_transforms(transforms,bs=8,nrow=2, to_inv=False):
train_data = Plant_Dataset("data/train_images",
train_df_cp,
transforms=transforms,
train=True)
# print(len(train_data))
loader_train = DataLoader(train_data,
batch_size=bs,
shuffle=False,
num_workers=4)
imgs = next(iter(loader_train))
plt.figure(figsize=(20,10))
batch = torchvision.utils.make_grid(imgs[0], nrow=nrow).permute(2, 1, 0).detach().cpu().numpy()
batch = cv2.cvtColor(batch, cv2.COLOR_BGR2RGB)
plt.imshow(batch)
plt.axis('off')
plt.show()
if to_inv:
return batch
In [ ]:
train_transform = A.Compose([
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.Rotate(
always_apply=True,
p=1.,
limit=(-68, 178),
interpolation=1,
border_mode=1,
# value=(0, 0, 0),
mask_value=None
),
ToTensorV2(),
])
show_transforms(train_transform)
Weather¶
In [ ]:
train_transform = A.Compose([
A.RandomShadow(
num_shadows_lower=1,
num_shadows_upper=3,
shadow_dimension=5,
p=1.,
always_apply=True
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.ShiftScaleRotate(
shift_limit=0.2,
scale_limit=0.2,
rotate_limit=90,
p=1.,
always_apply=True
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.RandomFog(
fog_coef_lower=0.2,
fog_coef_upper=0.5,
alpha_coef=0.3,
p=1.,
always_apply=True
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.RGBShift(
r_shift_limit=20,
g_shift_limit=20,
b_shift_limit=20,
always_apply=True,
p=1.
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.RandomBrightnessContrast(
always_apply=True,
p=1.
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.CLAHE(
always_apply=True,
p=1.
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.GaussNoise(
var_limit=(100, 170),
always_apply=True,
p=1.
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
train_transform = A.Compose([
A.CoarseDropout(
max_holes=5,
max_height=7,
max_width=7,
min_holes=3,
min_height=5,
min_width=5,
always_apply=True,
p=1.
),
ToTensorV2(),
])
show_transforms(train_transform)
In [ ]:
insect_transforms = Compose([
InsectAugmentation(insects=5, always_apply=True, p=1.),
ToTensorV2()])
show_transforms(insect_transforms)
In [ ]:
train_transform = A.Compose([
OneOf([
A.Rotate(
limit=(-68, 178),
interpolation=1,
border_mode=1,
mask_value=None),
A.ShiftScaleRotate(
shift_limit=0.2,
scale_limit=0.2,
rotate_limit=90)]),
OneOf([
InsectAugmentation(insects=5),
A.CoarseDropout(
max_holes=5,
max_height=7,
max_width=7,
min_holes=3,
min_height=5,
min_width=5),
A.RandomShadow(
num_shadows_lower=1,
num_shadows_upper=3,
shadow_dimension=5)
]),
OneOf([
A.RandomFog(
fog_coef_lower=0.2,
fog_coef_upper=0.5,
alpha_coef=0.3),
A.CLAHE(),
A.RandomBrightnessContrast(),
A.RGBShift(
r_shift_limit=20,
g_shift_limit=20,
b_shift_limit=20),
A.GaussNoise(
var_limit=(100, 170))]),
# A.Normalize(
# mean=(0.485, 0.456, 0.406),
# std=(0.229, 0.224, 0.225)),
ToTensorV2()])
show_transforms(train_transform, bs=32, nrow=4)
